ctx vs Gemma 4 Multimodal Fine-Tuner

ctx is an Agentic Development Environment (ADE) that solves the proliferation problem every developer hitting multi-agent workflows faces: you want to run Claude Code on one task, Codex on another, and Cursor on a third — but you end up with three terminal windows, three context streams, and no unified way to review what any of them did. ctx provides one controlled surface for all of them, with containerized disk and network isolation, durable transcripts, and a merge queue system that keeps parallel worktrees from colliding. The security model is where ctx gets interesting for teams. Platform and security teams get a single controlled runtime instead of hoping developers are running agents responsibly. Agents operate with bounded autonomy rather than requiring constant approval — you set the disk and network controls upfront, then let them run. All tasks, sessions, diffs, and artifacts land in one review surface you can search and audit. Shown on Hacker News today and currently free with an open-source GitHub repository (github.com/ctxrs/ctx), ctx is positioning itself as the layer between developers and their AI agents — the place where you actually manage what the agents are doing rather than just talking to them one at a time. With 23 supported CLI agents including Claude Code, Codex, Hermes Agent, and Amp, it's already broad enough to be genuinely useful.

Gemma 4 Multimodal Fine-Tuner is an open-source toolkit that lets developers fine-tune Google's Gemma 4 and 3n models across all three modalities — text, images, and audio — using only Apple Silicon hardware. It runs natively on PyTorch with Metal Performance Shaders (MPS), bypassing the NVIDIA requirement that has historically blocked Mac users from serious local fine-tuning work. The toolkit handles the full training pipeline including dataset prep, LoRA adapters, and multi-modal data collation. It ships with working example notebooks, a validation suite, and clean abstractions that don't require deep familiarity with the underlying MPS stack. Apple Silicon's unified memory architecture actually helps here — large multimodal batches fit in memory that would otherwise require GPU VRAM splitting on CUDA setups. Posted to Hacker News on April 7 as a Show HN, it pulled 109 upvotes and 165 GitHub stars within hours. The timing is sharp: Gemma 4 just dropped days ago with new multimodal capabilities, and the community immediately wanted local fine-tuning. This fills that gap faster than Google's own tooling.